Netherlands Institute for Radio Astronomy 2013. 7. 10. · CEP3 Hooking on to RUG Millipede...
Transcript of Netherlands Institute for Radio Astronomy 2013. 7. 10. · CEP3 Hooking on to RUG Millipede...
Netherlands Institute for Radio Astronomy
CEP3
July 10, 2013
CEP3: objectives
( Replacement for CEP1
( CEP1 hardware no longer supported & end of life
( Operational cost high (manpower & energy)
( Specification (budget constraints may impose deviations)
( 5 Tflop/s CPU power
( Memory/core ratio > 3 Gbps
( Workspace/core ratio > 100 GB
( 400 TB net storage (for data staging, not long term storage)
( 40 Gbps storage – compute IO
( Ready for future developments
( GPU capability
( Adding/access to special nodes (e.g. more memory)
( Possibility for scaling up
( Batch scheduling for efficient resource usage
CEP3 Hooking on to RUG Millipede replacement
( Current RUG Millipede (Groningen University HPC cluster since 2010)
( 236 nodes with
( 12 Opteron 2.6 GHz cores
( 24 GB memory
( 16 nodes with
( 24 Opteron 2.6 GHz cores
( 128 GB memory
( and 1 node with
( 64 cores
( 512 GB memory
( Network
( Infiniband switch
( Storage
( 110 TB
CEP3 Hooking on to RUG Millipede replacement
( RUG Millipede replacement
( Phase 1:
( After this Summer: ~ 64 servers added to existing HPC cluster
( Phase 2:
( End 2013/Early 2014: full replacement of existing hardware
( Server hardware
( 16 core Intel CPU’s
( 64 GB
( 2x 3TB internal disks
( Phase 1: ~ 16 fitted with NVIDIA Tesla K20M
( Storage
( Lustre
( ~500 TB net storage
( ~25 Gbps IO
CEP3 As a Millipede ‘add-on’
( Worker nodes
( Millipede nodes fulfil requirements
( For CEP3 initially CPU-only nodes (…)
( Option to share resources -> access to large & GPU capability Millipede queues
( Storage
( Millipede Lustre storage has its benefits...
( HPC ‘optimized’ (balancing, all nodes share same storage, scalability, etc)
( …and drawbacks
( Robust hardware architecture relatively expensive and limited IO
( For CEP3 looking at alternatives
( e.g. inexpensive ‘non-Lustre’ storage or integrate storage with worker nodes
( Possible Timeline
( Aim to decide on architecture/investment plan before August
( Acquisition September
( Available < 2014